The AI content engine that writes catalogue pages without flattening the brand

I built a thing this year that wrote 4,134 product descriptions for a single catalogue, and the number is the least interesting part of it. Because somewhere in the middle of that build—which I figured out the hard way, which is also why I trust it now—I worked out that the part I'd assumed was the achievement wasn't the point at all.

The obvious read is "look how much it wrote", and sure, 4,134 is a lot, and there's a real case for it—no human should spend their best thinking hours on the 1,307th variant of a connector description and pretend to enjoy it, and I've watched people try, and it doesn't end well for the connector or the human. But the volume was never the bit worth bragging about. The thing underneath it was.

Here's the trap, and I fell into it before I built my way out of it. Catalogue content looks simple right up until you're 600 products deep and the word premium has lost all legal meaning, and you realise that somewhere around the 200th description you stopped writing about the product and started writing a horoscope. And the bad version of AI makes that worse, not better—it writes fast, which everyone notices, and it writes as if every product got handed the same brief by the same committee, which takes longer to notice and a lot longer to undo. Everything effortless, everything elevated, everything designed to enhance your lifestyle, until the whole catalogue reads like a hotel lobby that's run out of things to say.

So the useful version was never "ask AI to write product descriptions" and hope. It's an engine, which is a duller word and a much better thing to own.

An engine has rails

The rails are the whole point, and they start with the inputs—the product facts, the category rules, the brand voice, the language your customers actually use, the SEO fields, the internal-linking logic, and, just as importantly, the short list of things the model is flatly not allowed to invent. Feed it that and it has somewhere to stand.

Then there's a format it fills every single time, the same shape for every product—a title, a short description, a long one, the metafields, FAQs where they earn their place, collection copy, internal links, alt text, whatever the store genuinely needs and nothing it doesn't. The same shape every time, on purpose, which sounds tedious and is exactly why it works.

And then there are the checks, which are the part most people skip and the part that, honestly, matters most—does the copy claim a feature the product doesn't have, does it match the product data, are the headings clean, do the links go anywhere, did it reach for one of the phrases you banned. You catch that before it ships, not after a customer does. With those rails in place you can write at catalogue scale and the catalogue still doesn't feel mass-produced, which is the whole reason to do it this way.

The voice has to come from somewhere real

Brand voice isn't a list of adjectives. "Warm, premium, playful, confident" gives the model a mood to perform, not a voice to write in, and the difference shows up by about the third paragraph.

The inputs that actually work are concrete and a bit unglamorous—past emails, the product pages the founder quietly likes, the reviews, the support replies, the phrases customers use without being prompted, the way the brand talks when it stops trying to sound like a brand and just says the thing. That last one does most of the work.

And this matters more than people think, because AI loves average language—it has read the internet, after all, so left without strong gravity it drifts toward whatever everyone else would have said. A good engine is mostly an exercise in giving it something heavier than the average to fall toward.

So if scale wasn't the prize, what was

Consistency. That's the answer I landed on. Every page carrying the right facts in the same structure, every SEO field present because the format demanded it and not because someone remembered, every internal link there for a reason, so the next range update is a normal Tuesday instead of starting the whole thing over. And because the pipeline is reusable, catalogue content stops being the once-a-year clean-up a business dreads and becomes a thing it just runs—new products go in, good pages come out, and someone still reviews the bits that carry taste or risk, because those stay human jobs. The grind comes off the human plate, which, after watching people grind, is the part I actually care about.

Where this helps most

• large catalogues with thin or inconsistent descriptions
• technical products where the facts actually matter
• brands with a strong voice but weak product-page discipline
• stores planning SEO growth across category and product pages
• teams that launch new products faster than they can write about them

If any of that sounds familiar, the problem probably isn't that you need a copywriter to grind harder—it's that the work needs shaping into something repeatable, so the catalogue stops quietly eating the time of the one person who knows the brand well enough to write it properly.